A MIMO-ANN System for Increasing Data
Rates in Organic Visible Light Communications
Systems
Paul Anthony Haigh1, Zabih Ghassemlooy1, Ioannis Papakonstantinou2, Francesco Arca3,
Sandro Francesco Tedde3, Oliver Hayden3 and Sujan Rajbhandari1
1
Optical Communications Research Group, Northumbria University, Newcastle-upon-Tyne, NE1 8ST
2
Department of Electrical & Electronic Engineering, University College London, London, UK
3
Siemens AG Corporate Technology, Erlangen, Germany
[email protected]
Abstract—This paper presents the first ever experimental
demonstration of a multiple-input multiple-output (MIMO)
visible light communications system employing four silicon (Si)
light emitting diodes (LEDs) and four organic photodetectors
(OPDs) as transmitters and receivers, respectively. The proposed
link is relatively low cost and it employs the on-off keying (OOK)
modulation format offering a data rate of 200 kb/s without the
need for equalization, which is a significant increase compared
with previous non-equalized systems. In order to speed up date
rates further, we implement an artificial neural network (ANN)
to classify the signal and correct the error induced by the matrix
inversion at the receiver, allowing a gross bit rate of 1.8 Mb/s in
the best case.
Index Terms— Artificial neural network, multiple-input
multiple-output,
organic
photodetector,
visible
light
communications
I. INTRODUCTION
M
ULTIPLE-input multiple-output (MIMO) systems
based on the concept of channel matrix inversion [1]
are an excellent way to increase the aggregate bit
rates in visible light communications (VLC) systems. As a
result they have garnered a significant amount of interest in
the research community with many different methods of
implementing the MIMO technique being proposed [1-6]. For
example, different modulation schemes are reported in [2] and
[5] while [4] offers a spatially diverse scheme where only one
transmitter is active at any given time. Furthermore, based on
the simulation results in [3] there is a strong claim that a
complex imaging receiver is required in order to receive an
error free link [1, 3, 7] due to an ill-conditioned channel
matrix rank.
In this paper we experimentally verify that it is possible to
recover a 4 × 4 MIMO link without employing a complex
imaging receiver at every point on the receiving plane while
also setting a world record bit rate of 1.8 Mb/s for a non
return-to-zero OOK VLC link that uses Si LEDs and OPDs.
While the benefits of solid state lighting in VLC are well
known (high output power, bandwidths of several MHz [8]),
using OPDs as the receiver for VLC systems is a subject in its’
infancy. So far only a single-input single-output (SISO) link
has been reported [9] where an artificial neural network
(ANN) equalizer is being employed to increase the data rate to
750 kb/s considering that the OPD bandwidth is only 30 kHz.
In this work we present for the first time a MIMO system
based on four individual commercial Si LEDs and four OPDs
that are mounted on the same substrate.
The OPD used is based on the bulk heterojunction (BHJ)
concept [10], which is an interpenetrated blend of electron
donor and electron acceptor, as opposed to the common p–n
junction that Si PDs utilizes. It was manufactured by Siemens
AG Corporate Technology by the spray deposition technique
as reported in [11]. When comparing organic technologies
with their crystalline Si counterparts, the real advantage is the
price. The total cost of materials for the poly(3hexylthiophene):[6,6]-phenyl C61-butyric acid methylester
(P3HT:PCBM) system implemented here is ~ € 0.20/cm2,
which is very inexpensive. The materials can be processed
into a solution that can be sprayed onto the substrate, allowing
very low cost and straightforward manufacturing. Besides
spray coating there are other ways to produce devices, see [12]
for further details on a number of methods. Additionally, the
OPD BW is dynamic [13], and its performance is controlled
by the incident light intensity. Under a high light intensity (>
300 μW/cm2) the number of charge carriers generated is
greater than the number of traps at the interface meaning that
the time constant of the plate capacitance controls the cut-off
frequency as in Si PDs. Conversely at low light intensities the
number of traps is greater than the number of charge carriers,
meaning that the BW is controlled by the time constant of the
interface traps.
A further advantage of the OPD under test is that is has four
independent diodes spaced 1.2 cm apart each with 1 cm2
photoactive area meaning that 4 × 4 (4-transmitters and 4receivers) MIMO is a natural progression on SISO links. No
additional electronics or photodetectors are required at the
receiver as in Si based MIMO, and therefore there is no
additional cost.
Since there are four diodes on the OPD substrate, 4 × 4
MIMO is investigated in this paper. While the theory of
demodulating a MIMO link is extremely simple (outlined
mathematically in Section II), a series of pilot tones are
transmitted from the first transmitter to all receivers in order to
find the channel response (while keeping the rest of the
transmitters off) and repeating for each transmitter. A matrix
Fig. 1 System block diagram. The transmission side is controlled by LabVIEW whereas the demodulation is
performed in MATLAB
of channel coefficients is produced, inverted and multiplied
with the received data in order to find the transmitted data. In
practice this is a significant challenge for band limited systems
due to the intersymbol interference (ISI) induced distortion of
the reference signals [9] that make up the channel matrix
coefficients.
MIMO systems are typically implemented in the radio
frequency domain where a multipath environment is inherent.
There are many reports of using an ANN to find the channel
response [14, 15]; however to the best of our knowledge there
are no reports of applying an ANN at all in VLC MIMO
systems. This is because they are mostly used for correcting
ISI at the receiver. One cause of ISI is transmitting data
outside of the modulation BW, which is not a problem that is
observed in MIMO systems explicitly, since the quality of the
channel coefficient matrix has more impact in recovering the
data than the ISI. In this paper we demonstrate that not only is
it possible to implement the ANN to equalize the received
signal but it is also possible to use the ANN to set a new
record data rate based upon the ANN for a large coverage area
beneath the transmitter array.
This paper is organized as follows. In Section II the MIMO
theory and test setup is outlined while in Section III the ANN
is described. Section IV shows the results and in Section V
conclusions are drawn.
II. MULTIPLE-INPUT MULTIPLE-OUTPUT
The concept of MIMO is outlined in Fig. 1. Over the next
few subsections the transmitter, channel and receiver are
outlined.
A. Transmitters
A pseudorandom binary sequence is generated and shaped
into an OOK signal in MATLAB. Two arbitrary function
generators (AFGs) with a peak-to-peak output voltage of 5 V
are synchronized. Synchronization of the AFGs is not required
as the data is recoverable regardless; however it does simplify
the demodulation process. Autocorrelation will only be
required once for all channels as the relative delay at the
receiver will be the same. Each channel is buffered using a
NAND gate with high output impedance then mixed with a
bias current of 350 mA. The bias current is generated by a
simple transistor circuit given in [16] to eschew the use of a
coupling capacitor and therefore avoid the baseline wander
phenomena that occurs in organic VLC links [9]. The
transmitter array is made up of four yellow phosphor blue chip
Philips Luxeon Rebel DS64 LEDs. The blue chip LED has a
high order modulation BW; however the slow response of the
yellowish phosphor limits the BW to several MHz. This BW is
still approximately an order of magnitude higher than that of
the OPD, which means that the OPD is limiting the link BW.
The LED is a Lambertian emitter as given by [17]:
௠ ାଵ
ܴ଴(ߠ) = ቂ ቃcos(ߠ)௠
(1)
ଶ
where m is the Lambertian number and θ is the angle of
emission. The transmitter array is divided into four serial data
streams, each transmitting a unique pseudorandom binary
sequence.
B. Channel matrix
At the receiver the incident light from each transmitter adds
by superposition to form the MIMO symbol. The VLC
channel gain from transmitter i to receiver j is given by [18]:
஺
ℎ௜௝ = మೝ ܴ଴൫ߠ௜௝൯cos൫߮ ௜௝൯
(2)
ௗ೔ೕ
where Ar is the photodetector area, dij is the distance between
transmitter i and receiver j, θij is the emission angle and finally
ϑij is the angle of incidence. Clearly, if there is no light
incident to the receiver, hij = 0. From (2) it is noteworthy that
there is no phase component, meaning that no phase distortion
occurs in the channel and only a flat attenuation is experienced
i.e. it is a DC gain < 1. There is a line of sight (LOS) path
between each transmitter and receiver; considering there are
four transmitters this gives a total of 16 LOS paths. It should
be noted that there is a multipath component;, however studies
show that the strongest multipath component is at least 7 dB
lower than the LOS component [3], so the multipath effect is
ignored in this experiment. Using 16 LOS paths a channel
matrix H containing all the useful information about the
channel can be built as [17]:
 h11
h
H   21
h31

h41
h12
h13
h22
h23
h32
h33
h42
h43
h14 
h24 
h34 

h44 
(3)
Hence, the received electrical signal vector immediately
after the photodetector is given by [1]:
‫۾‬௥௫ = ۶‫۾‬௧௫ + ‫ܖ‬
(4)
where ‫۾‬௧௫ and ‫۾‬௥௫ are transmit and receive vectors,
respectively of size 4 × η, where η is the number of
transmitted symbols and ‫ ܖ‬is the noise vector. Therefore the
transmitted symbols can be estimated by [1]:
ૐ = ۶ ିଵ‫۾‬௥௫
(5)
where ψ is the estimate of Ptx. Notice that for the recovery of
data, the channel matrix H must be full rank. It is clear that the
quality of the estimated data is dependent on H. The quality of
the H-matrix consequently depends on the quality of
measurement of the individual channel coefficients. The
channel coefficient hij is controlled by dij2 as (θij and φij change
as a function of modifying dij2) while Ar is a constant. A
further consideration is that ψ is dependent on the noise n. So
far, no studies have been carried out on the impact of the noise
on H, which is something we expect to do in future studies.
Tracking dij2 in real time to build the H-matrix would be very
problematic experimentally as it would involve a feedback
channel - something not implemented in this work.
A far more simplistic approach is a histogram method. One
LED transmits data over the channel while the remaining
receivers are not active. The DC level is removed from the
signal and the electrical signal levels are found. Recalling Vpp
= 5 V, removing the DC corresponds to signal levels of  2.5
V. Therefore it is easy to rapidly find the DC gain of the given
channel with a simple division. The method is then repeated
for each LED in the transmitter to build H – see Fig. 2 for a
two channel example.
In this test, the histogram measurement is only performed
once for the each channel because the receiver is not moving
continuously. For a real time measurement, H would be
periodically updated.
C. Receiver
The x – z and x – y plane geometries are illustrated in Fig.
3(a) and Fig. 3(b), respectively where Fig. 3(a) demonstrates
the previously outlined concepts and Fig. 3(b) shows the
receiver plane geometry. The distances between the x – z plane
and the x – y plane are both 10 cm. Only one quarter of the
possible area is tested, since the performance is expected to be
symmetrical around the center. This area is separated into nine
sections of 5 cm2 (S1:S9) to give an outline of the positiondependent performance. However before the performance is
analyzed, it is necessary to characterize the OPD under test.
z
LED
10 cm
OPD
x
(a)
y
5 cm
S1
S2
S3
S4
S5
S6
S7
S8
S9
5 cm
10 cm
10 cm
x
(b)
Fig. 2 A two channel example of the histogram method
exploited to find the channel response
There is a drawback to this method; when transmitting
outside the BW of the OPD, the influence of ISI will be
significant, degrading the peaks of the histogram meaning the
estimate of the signal will not be optimal. In the presence of
severe ISI the histogram will fail completely since the
histogram will not find any distinction between the two levels.
Fig. 3 (a) x – z plane (distance 10 cm) and (b) x – y plane
(distance 10 cm): the receiver plane divided into sections
S1-S9 for BER measurements
As mentioned, there are four diodes on the substrate, as
illustrated in Fig. 4, which also shows the spatial
characteristics of the OPD and the responsivity of the OPD is
shown in Fig. 5. The distance dij between the transmitter and
receiver is set to maximize the light density while maintaining
full coverage by all other LEDs. Within the receiving plane,
the maximum light density is found directly beneath one of the
LEDs [4], i.e. section S3, which corresponds to a light density
of > 300 μW/cm2 and a BW of 177 kHz. The minimum is
found in the center of the receiving plane; section S7 where
the light density is ~ 300 μW/cm2 and the BW is 133 kHz, see
Fig. 6.
The light density is measured with a thermopile of area 12
cm2 and the bandwidth is found by performing a fast Fourier
transform on a short pulse (duration 1 μs with rise and fall
times of 2 ns) transmitted over all 4-LEDs. The BW decreases
from a maximum of 177 kHz to 133 kHz as the receiver
approaches the center of the receiving plane. This difference is
quite large in relation to the magnitude of the BW. However,
in terms of a MIMO signal at the center of the receiving plane
the contribution from each LED is much greater, so it is
expected that this position will offer the largest data rate.
Normally it would be predicted that the largest BW would
offer the greatest data rate, yet since the bit rate will exceed
the BW in all sections, ISI will distort the signals significantly
thus causing the histogram method to fail.
Fig. 6 BW in the highest and lowest light densities on the
receiving plane
III. ARTIFICIAL NEURAL NETWORK EQUALIZER
Fig. 4 Bottom view photograph of the OPD showing the
spatial characteristics
Literature clearly demonstrates that the ANN is the most
effective method for equalizing ISI induced by a
communications channel [19-23]. The most popular ANN for
channel equalization is the single layer multilayer perceptron
(MLP). This is because it outperforms all other linear
transversal equalizers and can map any input-out sequence
including non-linear functions [19]. The ANN used in this test
has N input taps, N neurons and one output node γ. Each input
tap xi, where i = 1, …, N, has an associated weight wi. The
weight values are found by training with the LevenbergMarquardt back propagation (BP) algorithm that aims to
reduce the error function E(n) between a header sequence δ(n),
known at the receiver and the estimated data ψ(n) from the
MIMO demodulation [9]:
‫ )݊(߰‖ = )݊(ܧ‬− ߜ(݊)‖ଶ
(6)
The aforementioned weights between input i and neuron j are
adjusted as a function of E(n) such that [19]:
డா(௡)
‫ ݓ‬௜௝(݊ + 1) = ‫ ݓ‬௜௝(݊) − ߟ
(7)
డ௪ ೔ೕ(௡)
Fig. 5 OPD Responsivity in comparison to a Si PD – in the
visible region (~ 400 – 800 nm); the OPD has superior
responsivity in the visible range and blocks infrared
(meaning no additional infrared filtering is required) due
to a band gap energy of ~ 2 eV.
The data is transferred in each section and sampled with an
Agilent DSO9254A real time scope controlled by LabVIEW.
In the histogram method, data demodulation and estimation
are all performed in MATLAB. While a field programmable
gate array (FPGA) could have been implemented for a real
time result, there was a strong indication from the literature
that the experiment would not prove conclusive without an
imaging lens [1, 3, 7]. Therefore to avoid the costly
development time MATLAB is selected for this first
demonstration and we will proceed to FPGA development in
future work.
where η is the learning rate parameter, which if set too large
causes an unstable system and if set too low means slow
convergence. Therefore we use the adaptive converging rate
parameters to optimize convergence rates [24]. The output y is
given by [19]:
ߛ = ݂(ܾ + ∑௜‫ ݓ‬௜‫ݔ‬௜)
(8)
where i = 0, …, N if a bias b exists and i = 1, …, N otherwise.
While the ANN is normally used to correct the ISI induced
by a band limited system, in this case we use it to correct the
errors induced by an ill conditioned matrix inversion; i.e. the
ANN will not have any direct influence on the H-matrix, it
will aim to minimize the error function between the
transmitted and received data – the ANN is not performing
channel deconvolution.
IV. RESULTS
The Q-factors of the four channels in every position were
measured. The best results are found when the receiver is in
section S7 as predicted with the results as well as the eye
diagram shown in Fig. 7. A Q-factor of 4.7 is equivalent to a
bit error rate (BER) of 10-6, which is commonly accepted as
error free in VLC [9, 25]. Fig. 7 also depicts that a 200 kb/s
link can be established without an equalizer as two 100 kb/s
channels could be demodulated error free (Ch3 and Ch4). One
possible reason for this is due to the fact that the measurement
was made horizontally over the optical bench where the
transmitters and receivers are perpendicular to the bench, as
opposed to the case where transmitters and receivers are
parallel to the bench; thus resulting in large multipath
components for the two channels closest to the bench, i.e. Ch1
and Ch2 in this case. As mentioned, the rank of H is crucial
for recovering data; any loss of rank in the channel matrix
means that the data can’t be recovered. Since only two
channels could be demodulated error free, H is an illconditioned matrix and as a result a portion of the data is lost
as indicated by the condition number κ(H) = 70.
It should be noted that a 200 kb/s link with no equalizer is
significantly larger than the highest previously reported case
of 30 kb/s in [9] by around seven times. However a more
effective method of increasing data rates without an equalizer
might be to scale down the number of transmitters to two.
When considering the ANN equalizer, the achievable data
rate for each section with all channels error free is illustrated
in Fig. 8, where BER for all channels in sections S1, S3, S7
and S9 are averaged to produce a solitary curve rather than
four curves per section. No other curves are illustrated since
their performance lies between these sections. Sections S2 and
S6 can offer 1.5 Mb/s while sections S5, S5 and S8 can offer
1.7 Mb/s. As predicted section S7 offers the highest data rate
of 1.8 Mb/s while section S3 with the highest BW offers 1.4
Mb/s. This corresponds to four parallel channels transmitting
450 kb/s and 350 kb/s for each section, respectively. This
represents an extremely large increase of over 1 Mb/s on the
previously reported maximum data rate of 750 kb/s, not to
mention exceeding the modulation BW by an order of
magnitude. The performance difference of 300 kb/s observed
between S1 – S9 is not a function of the BW as previously
outlined, since the BW is maximized when the achievable data
rate is minimized. Therefore considering section S3, the
contribution to the MIMO symbol from the furthest LED is
minimized in comparison to section S7 where the relative
contributions from each LED are approximately equal
meaning there it is more likely that the symbol will be
recovered. This is reflected by the other sections, whereas the
deviation from the center point increases, the data rate
decreases.
Fig. 7 Received Q-factor for section S7 with eye diagram
inset at 50 kb/s The dashed line represents Q = 4.7,
corresponding to a BER of 10-6
It was suggested in [3] that section S7 would be impossible
to recover due to the matrix correlation, however the data
could recovered in these tests. While in theory, if the channel
paths and transmission power are identical, the matrix will
correlate [3]. In practice, however this situation is extremely
unlikely to occur due to the combination of the additive white
Gaussian noise (AGWN) that exists in the system plus the fact
that the LEDs are not perfectly identical so will have slightly
different output characteristics.
Fig. 8 Aggregate BER and bit rate for the four key sections
tested
In the worst case scenario, 1.4 Mb/s could be achieved,
which still denotes a sizeable improvement as mentioned. This
result is important not only for this reason, but because it
means that data can be demodulated at all positions on the
receiving plane indicating that MIMO is a valid technique for
future organic VLC systems. It is important not to forget that
in all cases an ANN was required to equalize the errors
induced throughout the system. Without the ANN, only 100
kb/s could be transmitted over two channels. While not close
to the 1.8 Mb/s reported above, in comparison to other
unequalized LED to OPD links, this is still an increase over
previously published work [9].
V. CONCLUSION
In this paper we have presented an experimental
demonstration of a MIMO system for an organic VLC system.
Whilst providing this demonstration, the record data rate of
750 kb/s for VLC systems with Si LEDs and OPDs as
transmitter and receiver, respectively has been well broken, as
1.8 Mb/s can now be transmitted, which is a significant step in
the progression of organic VLC. The major challenge in this
paper was to overcome the poorly conditioned matrix to
recover the data at the receiver. In order to achieve this, we
proposed an ANN equalizer to map the input-output sequence
and correct the errors, which resulted in achieving extremely
high data rates. We also demonstrated the feasibility of VLCMIMO without the use of an imaging receiver and for the first
time, multi-Mb/s an organic VLC link.
Future work on this topic considers what is missing from
this preliminary experiment. Firstly, an FPGA development is
required to provide a real time link where we can fully explore
the influence of white noise on H, which will allow the
derivation of new theory on the recovery of MIMO symbols
for VLC links. Secondly, in order to provide a full scale
demonstration, the number of OLEDs needs to be scaled up to
provide full illumination of a small room. It would be
expected that the link has the same performance as in the
small scale link that is presented here, and so far a full scale
demonstration has not been demonstrated.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
K. D. Dambul, D. C. O'Brien, and G. Faulkner, "Indoor Optical
Wireless MIMO System With an Imaging Receiver," Photonics
Technology Letters, IEEE, vol. 23, pp. 97-99, 2011.
A. H. Azhar, T.-A. Tran, and D. O'Brien, "Demonstration of highspeed data transmission using MIMO-OFDM visible light
communications," GLOBECOM Workshops (GC Wkshps), 2010
IEEE, pp. 1052 -1056, dec. 2010.
Z. Lubin, D. O'Brien, M. Hoa, G. Faulkner, L. Kyungwoo, J.
Daekwang, O. YunJe, and W. Eun Tae, "High data rate multiple
input multiple output (MIMO) optical wireless communications
using white led lighting," IEEE Journal on Selected Areas in
Communications, vol. 27, pp. 1654-1662, 2009.
R. Mesleh, R. Mehmood, H. Elgala, and H. Haas, "Indoor MIMO
Optical Wireless Communication Using Spatial Modulation," in
Communications (ICC), 2010 IEEE International Conference on,
2010, pp. 1-5.
T. Fath, M. Di Renzo, and H. Haas, "On the performance of Space
Shift Keying for optical wireless communications," in
GLOBECOM Workshops (GC Wkshps), 2010 IEEE, pp. 990-994.
Y. A. Alqudah and M. Kavehrad, "MIMO characterization of
indoor wireless optical link using a diffuse-transmisson
configuration," IEEE Transaction on communications, vol. 51, pp.
1554-1560, Sept. 2003.
A. H. Azhar, T. Tuan-Anh, and D. O'Brien, "Demonstration of
high-speed data transmission using MIMO-OFDM visible light
communications," in GLOBECOM Workshops (GC Wkshps), 2010
IEEE, 2011, pp. 1052-1056.
T. Komine and M. Nakagawa, "Fundamental analysis for visiblelight communication system using LED lights," Consumer
Electronics, IEEE Transactions on, vol. 50, pp. 100-107, 2004.
P. A. Haigh, Z. Ghassemlooy, H. Le Minh, S. Rajbhandari, F.
Arca, S. Tedde, O. Hayden, and I. Papakonstantinou, "Exploiting
Equalization Techniques for Improving Data Rates in Organic
Optoelectronic Devices for Visible Light Communications,"
Lightwave Technology, Journal of, vol. PP, pp. 1-1, 2012.
C. J. Brabec, N. S. Sariciftci, and J. C. Hummelen, "Plastic Solar
Cells," Adv. Funct. Mater., vol. 11, 2001.
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
S. F. Tedde, J. Kern, T. Sterzl, J. Fürst, P. Lugli, and O. Hayden,
"Fully Spray Coated Organic Photodiodes," Nano Lett., vol. 9, p.
980, 2009.
H. Klauk, Organic Electronics: Materials, Manufacturing, and
Applications: Wiley, 2006.
F. Arca, S. Tedde, B. Popescu, D. Popescu, S. Locci, M. Sramek,
P. Lugli, and O. Hayden, "Interface Trap States in Organic
Photodiodes," vol. [submitted], 2012.
K. K. Sarma and A. Mitra, "ANN based Rayleigh multipath fading
channel estimation of a MIMO-OFDM system," in Internet, 2009.
AH-ICI 2009. First Asian Himalayas International Conference on,
2009, pp. 1-5.
L. Zhang and X. Zhang, "MIMO channel estimation and
equalization using three-layer neural networks with feedback,"
Tsinghua Science and Technology, vol. 12, pp. 658-662, 2007.
A. R. Hayes, "Digital Pulse Interval Modulation for Indoor Optical
Wireless Communication Systems," PhD, Sheffield Hallam
University, UK, 2002.
J. R. Barry, J. M. Kahn, W. J. Krause, E. A. Lee, and D. G.
Messerschmitt, "Simulation of multipath impulse response for
indoor wireless optical channels," Selected Areas in
Communications, IEEE Journal on, vol. 11, pp. 367-379, 1993.
J. M. Kahn and J. R. Barry, "Wireless infrared communications,"
Proceedings of the IEEE, vol. 85, pp. 265-298, 1997.
K. Burse, R. N. Yadav, and S. C. Shrivastava, "Channel
Equalization Using Neural Networks: A Review," Systems, Man,
and Cybernetics, Part C: Applications and Reviews, IEEE
Transactions on, vol. 40, pp. 352-357, 2010.
C. Jongsoo, M. Bouchard, and T. H. Yeap, "Decision feedback
recurrent neural equalization with fast convergence rate," Neural
Networks, IEEE Transactions on, vol. 16, pp. 699-708, 2005.
F. Jiuchao, C. K. Tse, and F. C. M. Lau, "A neural-network-based
channel-equalization strategy for chaos-based communication
systems," Circuits and Systems I: Fundamental Theory and
Applications, IEEE Transactions on, vol. 50, pp. 954-957, 2003.
C. Jongsoo, A. C. C. Lima, and S. Haykin, "Kalman filter-trained
recurrent neural equalizers for time-varying channels,"
Communications, IEEE Transactions on, vol. 53, pp. 472-480,
2005.
X. Nan and L. Henry, "Blind equalization using a predictive radial
basis function neural network," Neural Networks, IEEE
Transactions on, vol. 16, pp. 709-720, 2005.
L. Behera, S. Kumar, and A. Patnaik, "On adaptive learning rate
that guarantees convergence in feedforward networks," IEEE
Transactions on Neural Networks, vol. 17, pp. 1116-1125, 2006.
J. Vucic, C. Kottke, S. Nerreter, K. Langer, and J. W. Walewski,
"513 Mbit/s Visible Light Communications Link Based on DMTModulation of a White LED," Lightwave Technology, Journal of,
vol. 28, pp. 3512-3518, 2011.